What is Cardinality-Balanced Multi-Target Multi-Bernoulli Filter?
Cardinality-Balanced Multi-Target Multi-Bernoulli Filter is a Random Finite Set approximation that represents a target population as independent Bernoulli components and corrects the positive cardinality bias of the original MeMBer update.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Represent existence separately for each object component
A Bernoulli RFS is empty with probability 1-r and contains one state drawn from p(x) with probability r. A multi-Bernoulli RFS is the union of independent components {(r_i, p_i)}. Its expected cardinality is the sum of existence probabilities, while its full cardinality distribution is obtained by convolving the Bernoulli probabilities.
This representation differs from CPHD's IID-cluster model. CBMeMBer keeps component-specific state densities and existence probabilities, but its independence approximation does not retain the full association dependence of a mixture posterior.
Correct the MeMBer update's cardinality bias
Vo, Vo, and Cantoni's CBMeMBer paper shows that the original MeMBer measurement update can positively bias target count and derives a cardinality-balanced multi-Bernoulli approximation under the same standard assumptions. Measurement-generated components use corrected existence terms rather than the biased MeMBer expression.
The correction is tied to the declared point-target, independent-detection, and clutter model. It does not make the posterior exact, and it should not be transferred unchanged to extended targets, correlated detections, or unknown clutter without a corresponding derivation.
Control approximation, extraction, and component growth
Gaussian-mixture implementations are convenient for linear Gaussian models, while Sequential Monte Carlo implementations support broader nonlinear or non-Gaussian dynamics. Both require pruning, merging, capping, and a declared extraction rule. Removing low-existence components also removes probability mass and can bias downstream counts.
Compare count RMSE and calibration, GOSPA or OSPA components, localization error, false and missed objects, component count, and latency across detection and clutter regimes. A tracking-and-classification implementation illustrates how CBMeMBer can carry additional class state, but its simulation results do not establish universal superiority.
Key Characteristics
- Represents each potential object by existence probability and state density
- Induces a Poisson-binomial target-cardinality distribution
- Corrects positive cardinality bias in the original MeMBer update
- Avoids emitting an explicit deterministic measurement assignment
- Supports Gaussian-mixture and particle implementations
- Does not preserve rigorous track identity without labels or trajectories
Common Use Cases
- Multi-target radar tracking with changing object count
- Particle-based tracking under nonlinear or non-Gaussian models
- Gaussian-mixture tracking under linear Gaussian assumptions
- Joint state and class estimation with object-specific components
- Baselining labeled RFS and PMBM trackers against an unlabeled filter
Example
Loading code...Frequently Asked Questions
What does CBMeMBer stand for?
CBMeMBer stands for Cardinality-Balanced Multi-Target Multi-Bernoulli. The capitalization preserves the historical MeMBer acronym. The method modifies the original MeMBer measurement update to remove a positive bias in its estimated number of targets.
How does a multi-Bernoulli filter represent target count?
Each independent component exists with probability r_i. The total count is the sum of those Bernoulli existence variables, so its distribution is Poisson-binomial and its expectation is the sum of all r_i. Thresholding components is an extraction decision, not the distribution itself.
How is CBMeMBer different from CPHD?
CPHD uses an IID-cluster approximation with one shared spatial density and an arbitrary cardinality distribution. CBMeMBer keeps a separate existence probability and state density for each component, but independence between components constrains the represented posterior.
Does CBMeMBer provide target identities?
No. An implementation may carry component indices or connect estimates between scans, but the standard unlabeled multi-Bernoulli posterior does not make those indices persistent identities. LMB, GLMB, or trajectory-set filters model identity more explicitly.
Which implementation choices most affect CBMeMBer results?
Birth components, detection and clutter models, proposal distributions, pruning and merging thresholds, component caps, and extraction thresholds all affect results. Report these settings with count and set errors, calibration, runtime, and failure rates.